0

I am using matlab's hmmtrain() to produce the final state transitions and emission probabilities matrices.

For the training I use the Baum-Welch algorithm, giving as input the sequences dataset, and initial guesses for A and B matrix.

Nevertheless, due to data sparsity I receive many probabilities equal to zero that is why I need to perform smoothing e.g. Laplace smoothing or other.

I read that there is hmmtrain(...,'Pseudoemissions',PSEUDOE) which applies only to Viterbi training algorithm. I thought that with this parameter I could achieve Laplace smoothing, but not sure if I can apply it to a "Baum Welch training" case. Second I am not sure how I should apply it. In the case of Laplace or add-one smoothing, should I add the positive number 1 to every cell of the PSEUDOE matrix or just where I expect to receive zero probabilities?

Thank you in advance for your help.

user2008973
  • 435
  • 3
  • 9
  • 22
  • 2
    What do you expect of an answer? Obviously, you can. What have you tried, what's your approach, how are things stored? – Marcus Müller Feb 10 '15 at 12:56
  • Ok perhaps a first kind of help would be an hmm matlab link that refers to smoothing? Because I haven't found anything relevant. Is there anything built-in? – user2008973 Feb 10 '15 at 13:10
  • It's matlab. It has *anything* ;) no, seriously, define "smoothing" mathematically, and search the matlab docs for what matches your definition – Marcus Müller Feb 10 '15 at 13:17
  • Thank you. Apparently I haven't asked in a clear way to get an answer that would cover my needs. I will come back if I have something more to add to my approach, but sincerely this my approach. I have applied hmmtrain() on my dataset, but in the matrices I receive zero probabilities for some observations. I do not know how to apply smoothing on my dataset. That's it, honestly I do not know what else I should provide as details. – user2008973 Feb 10 '15 at 13:22
  • @MarcusMüller I have updated my question providing more details. I hope now it is more comprehensible what I am looking for. Thank you again for any feedback you can provide. – user2008973 Feb 11 '15 at 11:45
  • Do I need to provide more details? – user2008973 Feb 17 '15 at 09:43

0 Answers0